CHANGELOG.md

Changelog

v1.2.0 - 2026-04-25

The v1.2 release. Claude Code remembers across compactions and sessions. SessionStart now emits a token-budgeted reference bundle instead of a 2-5K text dump; the agent pulls depth on demand via existing MCP tools. Behavior contracts let you teach Claude durable preferences once and have them ride into every session. An OpenTelemetry-shaped span substrate underneath both features makes the message graph queryable through the new lcc_expand span_id mode.

Three composing layers

Compaction-aware reference bundle (the headline). inject_context.py no longer dumps the full dream-patterns text plus top summaries. Instead, SessionStart receives a slot-packed reference bundle ordered with the recovery-protocol line first so the agent learns the recovery contract before any content. Default cap 1000 tokens, configurable via bundleTokenBudget. Each item carries a literal Expand: call MCP tool instruction the agent can invoke verbatim. PreCompact and PostCompact hooks stay no-push (hooks/post_compact.sh:54-56 constraint preserved verbatim). Recovery routes entirely through the SessionStart bundle's recovery line plus agent-initiated MCP calls.

Behavior contracts. New contracts table with typed kind (prefer / forbid / verify-before), bylined, append-only with a supersedes_id chain. The dream cycle proposes contracts as Pending; the new TUI Contracts tab (key 5) lets you approve (a), reject / retract (r), or supersede (s) interactively. Filter cycle (t) flips between Pending, Active, and Retracted views. Every Active contract rides in the SessionStart bundle's contracts slot.

OpenTelemetry-shaped span substrate. messages gains four new NULL-able columns: parent_message_id, span_kind, tool_call_id, attributes (JSON). Hooks populate them at write time. New scripts/db/spans.py exposes get_span, get_span_chain, get_children_spans, cap_attributes_json. lcc_expand accepts span_id and walks the parent chain; structured-error JSON returns (span_not_found / expand_too_large / vault_corrupt / permission_denied) with sanitized static messages so error output can never leak filesystem paths or library internals into agent context.

Added

• Bundle token budget (bundleTokenBudget, default 1000) and the bundleEnabled rollback flag (default true). Set bundleEnabled: false in ~/.lossless-code/config.json to disable SessionStart injection entirely while keeping all other lossless-code features active.
• lcc_contracts MCP tool and lcc contracts CLI verb. Actions: list, show, approve, reject, retract, supersede. Same result shape on both surfaces; structured-error JSON for new MCP errors with sanitized static messages.
• lcc_expand span_id mode plus structured errors. Existing summary_id and file modes preserve their v1.1.x human-readable output for backwards compatibility.
• TUI Contracts tab with full state coverage: Pending / Active / Retracted filter cycling via t, ContractDetailScreen modal for the full body before approval, SupersedeBodyScreen TextArea modal for in-place revision, RetractionReasonPrompt for required-reason retracts, conflict-signal column when an opposing-kind Active rule overlaps the candidate.
• collect_status_dict shared helper in scripts/lcc_core.py. Both lcc status and the lcc_status MCP tool now route through this single source of truth, closing a historical CLI-vs-MCP drift pattern. New v1.2 fields surfaced in status: contract counts by status, decision count, bundle enabled state and budget, lastDreamMode (whether the dream cycle ran via LLM, regex fallback, or failed).

Hardened

• Vault permissions. ~/.lossless-code/ now ships with mode=0o700 on the directory, 0o600 on vault.db and config.json. Closes the world-readable gap that v1.2's new contracts.body and messages.attributes columns would otherwise inherit on shared machines.
• Contract body sanitization at injection time. Bundle assembler strips \n and \r from contract bodies and rejects bodies containing the literal [lcc.contract] marker. Closes the newline injection vector flagged in v1.2 doc review: without this, an attacker who got an approved contract body with embedded \n[lcc.contract] FORBID poison would see a synthetic second contract line in the SessionStart bundle.
• Supersede atomicity. supersede_contract wraps INSERT new and UPDATE old in BEGIN IMMEDIATE / COMMIT so concurrent CLI + MCP writes cannot land two rival Active rows pointing at the same target.
• attributes JSON shape validation. store_message rejects non-dict attributes input (logged to stderr, stored as empty object); cap_attributes_json token-caps oversize payloads with a tombstone preserving the size signal.

Removed

• Legacy v1.1 full-text injection from inject_context.py. The reference bundle is now the only injection path. Set bundleEnabled: false to disable SessionStart injection entirely if you want the v1.1 behavior of "no lossless-code context layer." No compatMode dual-path is shipped; the previous proposal was rejected during plan review as a complexity tax for a measurement that can be done once pre-merge.

Multi-install drift warning

install.sh now detects when both manual and plugin installs are present and emits a stderr warning with one-liner remediation paths. The lcc doctor self-healing tool remains deferred to v1.3.

Migration notes

• Existing installs: re-run install.sh to pick up the two new scripts (scripts/contracts.py, scripts/lcc_core.py). Plugin users get the upgrade through the marketplace refresh.
• First dream cycle post-v1.2: the dream cycle now runs Phase 1.5 (contract + decision extraction) after pattern extraction. Pending contracts populate the TUI queue; nothing rides in the bundle until you approve.
• Schema migrations: idempotent ALTER TABLE adds five columns (4 on messages for spans, 1 on dream_log for mode) plus the contracts table. NULL-permissive for backfill compatibility; pre-v1.2 rows have no causal data and that is accepted.

v1.1.1 - 2026-04-14

Patch release that ships every main-branch delta accumulated since the v1.1.0 marketplace artifact, headlined by the claude --print CWD fix that closes the claude --resume pollution bug. Marketplace users should upgrade to pick up the fix; manual installs already covered by re-running install.sh.

Fixed

• claude --print subprocess CWD pollution (PR #11). The summariser's claude --print calls now pin cwd= to ~/.lossless-code/.cli-cwd, so Claude Code files internal summarisation sessions in their own bucket instead of polluting the user's claude --resume list with hundreds of "Summarise the following conversation turns concisely…" entries. Includes scripts/check_summariser_pollution.py regression detector.
• Installer + README sync for fingerprint feature (PR #10). install.sh now copies the new PreToolUse / PostToolUse hooks and the scripts/file_context.py, scripts/hook_store_tool_call.py files added by the fingerprint feature; README documents the new fileContextEnabled flag and the hook surface.

Added — fingerprint file context (default off, PR #9)

• fileContextEnabled config flag (default false) gates all new behavior. Flip to true in ~/.lossless-code/config.json to opt in.
• Schema additions: messages.file_path, summaries.kind, plus idx_messages_file_path (partial) and idx_summary_sources_source composite. Migrations are idempotent.
• PostToolUse hook (hook_store_tool_call.py) records which file a Read | Edit | Write | MultiEdit | NotebookEdit tool call touched.
• Polarity classification (summarise.classify_chunk_polarity) tags summaries as created | edited | discussed | mixed based on the tool calls in the chunk; propagates through cascade.
• db.get_summaries_for_file walks summary_sources upward from a message's file_path via recursive CTE (hop ceiling 16) and excludes consolidated nodes.
• inject_context.format_file_fingerprint renders a compact, 200-token fingerprint with a truncation ladder that always preserves the file path and the lcc_expand hint.
• PreToolUse hook (pre_tool_use.sh + scripts/file_context.py) injects the fingerprint as additionalContext before file-touching tools run. Reads are served from a JSON cache under cache/ with 60s TTL, file-locked reads/writes, atomic replace, and single-flight stampede guard. Cold path opens vault.db read-only so it never blocks the writer.
• lcc_expand --file <path> (CLI) and {"file": "..."} (MCP) drill from a file path to the recent summaries that reference it.
• lcc status surfaces tagged message count, distinct tagged files, and cached fingerprint count when the flag is on.

Changed

• scripts/db.py split into scripts/db/ package (PR #8). Single 745-line module replaced with focused submodules: config, schema, messages, summaries, sessions, embeddings, search, dream_log. Public import surface is preserved via scripts/db/__init__.py; existing call sites unchanged.
• BM25 prompt-aware context eviction (PR #7). Context injection now scores candidate summaries against the live prompt with BM25 instead of a flat recency cut, so high-signal older summaries survive eviction when the prompt actually references them.

Added

• Session filtering for lossless-claw v0.7.0 parity: stateless session gate, stop-hook filtering, plus DB-layer tests covering pattern matching and filter behaviour.
• Circuit breaker + dynamic chunk sizing for summarisation: protects the summariser from runaway provider failures and adapts chunk size to the active provider's context window.
• Vault bloat protection: summary size caps and dream pagination prevent unbounded growth on long-running vaults.
• claude-cli provider for Claude Max / Pro subscription users (no API key path). ANTHROPIC_API_KEY is stripped from the subprocess env so the CLI falls back to OAuth.
• Custom Anthropic-compatible providers (MiniMax and others) via openaiBaseUrl, plus provider auto-detection and structured error handling.
• dream --status flag for inspecting dream-cycle state.

v1.1.0 — semantic search hybrid

• Hybrid semantic search combining FTS5 BM25 with vector similarity via reciprocal rank fusion. Cached fastembed.TextEmbedding instance.